Detecting Temporally Localized Manipulations in Authentic Video Streams

Authors: Okan Umur, Ali Emre Güşlü, Ibrahim Delibasoglu

Published: 2026-06-05 09:35:51+00:00

AI Summary

This paper addresses the gap in existing datasets for detecting short, realistic manipulated segments embedded within otherwise authentic video streams. It introduces a custom-curated test set for this challenging scenario and evaluates two DINOv3-based approaches: a supervised linear probe and an unsupervised temporal feature similarity method. The study establishes an initial benchmark for partially manipulated video detection, highlighting the need for content-adaptive thresholding.

Abstract

The rapid advancement of video editing and generative artificial intelligence technologies has made realistic video manipulation increasingly accessible. Although existing datasets have significantly advanced research in deepfake detection, object removal, and video inpainting, they do not adequately model scenarios in which a short manipulated segment is inserted into an otherwise authentic video and the original video continues afterward. In this study, we review representative datasets from the literature, analyze their characteristics, and discuss their limitations with respect to temporally localized realistic manipulation detection. Based on this analysis, we motivate the need for a new dataset specifically designed for authentic videos containing short and highly realistic manipulated intervals. Finally, we evaluate two complementary approaches on our custom-curated test set to establish an initial benchmark for this challenging scenario. The first employs a linear probe on DINOv3 features, assessed under three thresholding strategies. The second leverages DINOv3 features with a consecutive frame similarity-based method to detect temporal manipulation boundaries. Together, these experiments provide an initial benchmark for partially manipulated video detection and highlight the need for content-adaptive thresholding mechanisms. The dataset, code, and supplementary materials are publicly available at https://github.com/OkanUmur/temporally-localized-video-manipulation-detection.

Key findings

The supervised linear probe struggles with cross-domain generalization, achieving modest precision (0.473 with adaptive threshold) but low recall. In contrast, the unsupervised DINOv3 feature similarity approach achieves a high global precision of 83.00% and exceptional video-level accuracy of 95.00% on the authentic control group. Both methods underscore the critical need for content-adaptive thresholding mechanisms to improve robustness and reduce false alarms.

Approach

The authors evaluate two complementary approaches. The first is a supervised linear probe trained on DINOv3 (ViT-Base) features, classifying frames as real or fake, and assessed with three thresholding strategies. The second is an unsupervised, training-free method that leverages DINOv3 features to compute cosine similarity between consecutive frames, using a rolling window Z-score and fixed thresholds to identify temporal manipulation boundaries.

Datasets

Custom-curated test set ('Pure Authentic Control Group' and 'Manipulated (Merged) Set'), DALL-E Recognition Dataset, publicly available stock footage repositories, and publicly available face datasets (for training the linear probe).

Model(s)

DINOv3 (specifically, ViT-Base variant) for feature extraction, and a linear classifier (linear probe).

Author countries

T¨urkiye

← Previous